ResNet-50 Network Enhancements: Improved Accuracy with Minimal Time Complexity Increase

Table 4 shows the final accuracy, recall, and single-class accuracy of three network models trained on the same dataset and training strategy. Figure 7 displays the validation loss of the three networks after 50 epochs, which stabilized at around 40 epochs. Both the error and accuracy metrics indicate that each proposed improvement has a positive impact on the classification task. The ResNet-50-a model, achieved by adding a CBAM module and modifying the downsampling method, outperforms the original ResNet-50 model by 2.50% in accuracy, with improvements in recall and single-class accuracy. The ResNet-50-b model, with further improvements, surpasses the ResNet-50-a model by 1.92% in accuracy and the original ResNet-50 model by 4.62%, achieving the highest recall and single-class accuracy. As the complexity of the model structure and the number of parameters increase from the original ResNet-50 to ResNet-50-a and ResNet-50-b, the time complexity of ResNet-50-a increases by 32.6%, while its accuracy improves by 2.50%. ResNet-50-b achieves a 4.62% improvement in accuracy with a 40.5% increase in time complexity compared to the original ResNet-50 model. These results demonstrate that the proposed improvements can significantly enhance the network's classification performance while only slightly increasing its time complexity.